Research Of Adaptive Facial Landmark Localization In-the-Wild

Facial feature point localization is an important part of face application,which plays an important role in facial pose correction,pose recognition,expression recognition,mouth recognition and other fields.However,facial images captured in natural scenes are often affected by uncontrollable factors such as illumination,posture,expression,occlusion and low image resolution,which bring great challenge to facial feature point localization.Therefore,how to locate facial feature points accurately in natural scenes is a key research issue in computer vision.Based on deep learning,this paper proposes an improved boundary-aware face alignment method by using stacked dense U-Net framework.The main work and achievements of this paper are as follows:(1)In order to build the algorithm framework of facial feature point localization based on heat-map regression,this paper discusses the basic knowledge of facial feature point localization.In the aspect of feature extraction,three kinds of symmetrical U-Net network structures are analyzed to extract the features of each scale of facial image.In the aspect of algorithm framework,the basic structure of convolution neural network and the realization mechanism of heat-map regression method are studied.In the facial feature point localization dataset,the common public dataset and the evaluation index of comparative experiment are summarized.(2)Regarding the representation of facial geometric information during face alignment,most methods directly use discrete feature points as facial geometric information representation.Based on the facial boundary information,this paper proposes an improved boundary-aware facial feature point localization algorithm.The network structure of the algorithm includes boundary heat-map estimation and facial feature point regression.In the boundary heat-map estimation stage,this paper divides the face according to the boundary,and the obtained face boundary information can effectively deal with occlusion and other situations,in which information is exchanged between the boundaries through the information transfer layer.In the regression stage of facial feature points,the heat-map of the face boundary and the original face image are fused,input into the stacked dense U-Net network,and deformable convolution is used for geometric transformation.The final training modelis experimentally analyzed on three datasets of 300 W,WFLW and COFW.The results show that the algorithm has high accuracy and is robust to facial images with large pose variation,expressions and occlusions.(3)Aiming at the problem of limited training dataset in model training,this paper uses a style translation data enhancement method.Stylized data enhancement decomposes the facial image into the style space of illumination,texture and image environment,as well as the face structure space with the same style.By using the method to separate and recombine the style space and structure space of the facial images,much more synthetic images are obtained to enhance the training data.Finally,the WFLW and 300 W benchmark datasets are enhanced by style translation,and the extended data sets are used for model training.Finally,the training results of the enhanced boundary aware feature point localization model are much better than the original training results.